Lumen

Inspiration

We wanted to tackle the problems of catching outliers/ anomalous data points. In this case, catching unlogged or misreported collections in real time. Accurate tracking matters, and we wanted a system that detects discrepancies and potential fraud as they happen.

What it does

CauldronWatch is a real-time monitoring dashboard that tracks potion levels across all cauldrons, detects collection events, and validates transport tickets. It:

Detects drain events using rate-of-change analysis and accounts for continuous filling during drainage
Matches tickets to actual drains and flags discrepancies (under-reporting, missing tickets, over-reporting)
Ranks cauldrons by overflow risk using travel times and capacity limits
Scores courier trustworthiness based on ticket accuracy
Visualizes the network with an interactive map showing cauldron locations, connections, and risk levels

How we built it

Backend: Python with pandas/numpy for time-series analysis and scipy for linear regression on filling segments
Detection algorithm: Analyzes level changes to detect drains, accounting for filling during drainage, and calculates volumes
Ticket matching: Matches drains to tickets by date/cauldron with configurable tolerance
Frontend: Streamlit dashboard with Plotly for interactive visualizations
Data pipeline: Fetches data from the API, processes it, detects anomalies, and surfaces insights

Challenges we ran into

Fill rate calculation: Median point-to-point rates were noisy. We switched to linear regression on long-term filling segments and used percentiles to capture steady-state rates
Drain detection accuracy: Accurate volume requires accounting for continuous filling. We track level drops plus the potion generated during the drain period
Matching logic: Tickets arrive at end of day with only dates, so we account for travel time to match drains correctly
Real-time processing: Optimized segment identification and regression to handle large time-series efficiently

Accomplishments that we're proud of

Detected 141 drain events with volume calculations
Identified 975L of unaccounted potion across multiple cauldrons
Built a trust scorecard that flags couriers with consistent under-reporting
Created an overflow prediction system that accounts for travel time and unload time
Developed a clean, modular codebase that's easy to extend and maintain

What we learned

Time-series analysis: Using linear regression on segments provides more accurate rates than point-to-point differences
Real-world constraints matter: Accounting for travel time, unload time, and continuous filling improves accuracy
Visualization helps: Interactive maps and charts make it easier to spot patterns and anomalies
Conservative estimates help: Using percentiles instead of medians provides more reliable steady-state measurements

What's next for CauldronWatch

Route optimization: Use the network graph to compute optimal courier routes to prevent overflow
Forecasting: Predict future levels and flag at-risk cauldrons earlier
Real-time alerts: Notify managers when discrepancies exceed thresholds
Historical analysis: Compare trends over time to identify seasonal patterns
Mobile app: Let couriers log collections on the go with automatic validation